Verdaca puts your question through a three-role loop: one model drafts, a second argues the other side and critiques the draft, and a third returns an answer, asks you to clarify, abstains or escalates. Every run writes a hashed receipt.
WIP · CHANNEL ACCESS
Planned: start a deliberation from a Microsoft Teams or Slack message. Today the gateway and the deliberation loop are separate; the channel adapters are tested against recorded traffic, not live workspaces.
MICROSOFT TEAMS·SLACK
The gateway identity-checks, replay-guards and budget-gates every inbound message. In tests, identity is verified against a local OIDC validator; one manual Auth0 run in June 2026, with no committed log.
WIP · HORIZONTAL · PLUG-IN PERSONAS
Planned: swap UX expert ↔ data scientist ↔ legal counsel ↔ security architect without changing workflow logic. Today the roles are fixed: producer, reviewer, synthesizer.
The persona is the argument. Change who's in the room and the answer changes — without touching a line of workflow logic.
VERTICAL · SUBSTRATE-NEUTRAL
Mem0 ↔ Letta (memory) and LLMLingua ↔ in-tree (compaction): proven by shared contract suites. OpenAI ↔ Anthropic and Forge ↔ Caveman: WIP, one adapter each, not swap-tested.
WHAT IT DOES
MULTI-AGENT DELIBERATION
A producer drafts; a reviewer writes a counterargument from the question and evidence alone, then critiques the draft; a synthesizer returns a typed outcome: answer, clarify, abstain or escalate. Errors fail closed to escalate.
ADAPTER REGISTRY
Every external dependency sits behind a ratified port contract. Swap the backing implementation without touching agent or workflow code: proven today for the memory store and the compaction engine; the model provider swap is WIP.
MEMORY SUBSTRATE
Mem0 (hybrid vector-plus-graph retrieval) and Letta (stateful session memory) both pass the same memory contract suite; either can back the memory port.
WIP · WORKFLOW RECIPES
Planned: pre-built deliberation patterns for strategy review, code audit, requirements analysis and UX critique. Design preview; not built yet.
RECEIPTS
Each run writes a JSON receipt: evidence, draft, critique, diff, outcome, and per-call model, tokens and cost. Hashed (SHA-256), not signed. Three recorded runs on one question, each about $0.03, are public.
PINNED AND CONTRACT-TESTED
Adapters are version-pinned and run against no-waiver contract suites in CI: 440+ root and 1,450+ kernel tests.
WORKING WITH US
Verdaca is early: the deliberation loop runs from a script or as an MCP tool (called once from Claude Desktop); Studio is WIP. We're working hands-on with a small number of design-partner teams before we set public pricing — shaping the product around real, high-stakes decisions. If your team has a decision worth deliberating, we'd like to talk.
SOURCED, VETTED, PINNED